-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathrun_epoch.py
180 lines (152 loc) · 7.75 KB
/
run_epoch.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
import torch, torch.nn as nn, torch.nn.functional as F
from pdb import set_trace as stop
from tqdm import tqdm
from utils.util import get_class
from models import *
from utils.loss import MCBCELoss
def run_epoch(args, model, data, optimizer, epoch, desc, device, loss_weight=None, train=False, warm=False, inference_with_sampling=False, stage='joint'):
if train:
model.train()
if isinstance(model, (ProbCBM)):
if warm and hasattr(model, 'cnn_module'):
for p in model.cnn_module.parameters():
p.requires_grad = False
elif hasattr(model, 'cnn_module'):
for p in model.cnn_module.parameters():
p.requires_grad = True
optimizer.zero_grad()
else:
model.eval()
# pre-allocate full prediction and target tensors
all_predictions = torch.zeros(len(data.dataset) * 2, args.num_labels).cpu()
all_certainties = torch.zeros(len(data.dataset) * 2, args.num_concepts).cpu()
all_cls_certainties = torch.zeros(len(data.dataset) * 2).cpu()
all_targets = torch.zeros(len(data.dataset) * 2, args.num_labels).cpu()
batch_idx = 0
end_idx = 0
loss_tot_dict = {'total': 0}
# Set criterion for class and concept
criterion_class = getattr(args, 'criterion_class', 'ce')
if criterion_class == 'ce':
criterion_class = nn.CrossEntropyLoss()
else:
raise ValueError('Got criterion_class', criterion_class)
criterion_concept = getattr(args, 'criterion_concept', 'bce')
if criterion_concept == 'bce':
criterion_concept = nn.BCEWithLogitsLoss()
elif criterion_concept == 'bce_prob':
criterion_concept = nn.BCELoss()
elif criterion_concept == 'MCBCELoss':
in_criterion = nn.BCELoss(reduction='none')
criterion_concept = get_class(criterion_concept, 'utils.loss')(criterion=in_criterion, reduction='mean', vib_beta=args.vib_beta, \
group2concept=args.group2concept)
else:
raise ValueError('Got criterion_concept', criterion_concept)
for batch in tqdm(data, mininterval=0.5, desc=desc, leave=False, ncols=50):
images = batch['image'].float().to(device)
target_class = batch['class_label'][:, 0].long().to(device)
target_concept = batch['concept_label'].float().to(device)
if train:
preds_dict, losses_dict = model(images, target_concept=target_concept, target_class=target_class, T=args.n_samples_train, stage=stage)
else:
with torch.no_grad():
preds_dict, losses_dict = model(images, target_concept=target_concept, target_class=target_class, inference_with_sampling=inference_with_sampling, T=args.n_samples_inference)
B = images.shape[0]
class_label_onehot, concept_labels, labels, concept_uncertainty, class_uncertainty = None, None, None, None, None
if args.pred_class:
class_labels = batch['class_label'].float()
class_label_onehot = torch.zeros(class_labels.size(0), args.num_classes)
class_label_onehot.scatter_(1, class_labels.long(), 1)
labels = class_label_onehot
concept_labels = batch['concept_label'].float()
if args.pred_concept:
if labels is not None:
labels = torch.cat((concept_labels, labels), 1)
else:
labels = concept_labels
assert (labels is not None)
loss, pred = 0, None
loss_iter_dict = {}
if args.pred_concept:
if isinstance(criterion_concept, MCBCELoss):
pred_concept = preds_dict['pred_concept_prob']
loss_concept, concept_loss_dict = criterion_concept(\
probs=preds_dict['pred_concept_prob'],
image_mean=preds_dict['pred_mean'], image_logsigma=preds_dict['pred_logsigma'],
concept_labels=target_concept, negative_scale=preds_dict['negative_scale'], shift=preds_dict['shift'])
if 'pred_concept_uncertainty' in preds_dict.keys():
concept_uncertainty = preds_dict['pred_concept_uncertainty']
for k, v in concept_loss_dict.items():
if k != 'loss':
loss_iter_dict['pcme_' + k] = v
elif isinstance(criterion_concept, (nn.BCELoss)):
pred_concept = preds_dict['pred_concept_prob']
loss_concept = criterion_concept(pred_concept, target_concept)
if 'pred_concept_uncertainty' in preds_dict.keys():
concept_uncertainty = preds_dict['pred_concept_uncertainty']
else:
pred_concept = preds_dict['pred_concept_logit']
loss_concept = criterion_concept(pred_concept, target_concept)
pred_concept = torch.sigmoid(pred_concept)
if 'pred_concept_uncertainty' in preds_dict.keys():
concept_uncertainty = preds_dict['pred_concept_uncertainty']
if stage != 'class':
loss += loss_concept * loss_weight['concept']
pred = pred_concept
loss_iter_dict['concept'] = loss_concept
if args.pred_class:
if 'pred_class_logit' in preds_dict.keys():
pred_class = preds_dict['pred_class_logit']
loss_class = criterion_class(pred_class, target_class)
pred_class = F.softmax(pred_class, dim=-1)
else:
assert 'pred_class_prob' in preds_dict.keys()
pred_class = preds_dict['pred_class_prob']
loss_class = F.nll_loss(pred_class.log(), target_class, reduction='mean')
loss_iter_dict['class'] = loss_class
if stage != 'concept':
loss += loss_class * loss_weight['class']
pred = pred_class if pred is None else torch.cat((pred_concept, pred_class), dim=1)
if 'pred_class_uncertainty' in preds_dict.keys():
class_uncertainty = preds_dict['pred_class_uncertainty']
for k, v in losses_dict.items():
loss_iter_dict[k] = v
if k in loss_weight.keys() and loss_weight[k] != 0:
loss += v * loss_weight[k]
loss_out = loss
for k, v in loss_iter_dict.items():
if v != v:
print(k, v)
if train:
loss_out.backward()
# Grad Accumulation
if ((batch_idx + 1) % args.grad_ac_steps == 0):
torch.nn.utils.clip_grad_norm_(model.parameters(), args.clip_grad_max_norm)
optimizer.step()
optimizer.zero_grad()
## Updates ##
loss_tot_dict['total'] += loss_out.item()
for k, v in loss_iter_dict.items():
if k not in loss_tot_dict.keys():
try:
loss_tot_dict[k] = v.item()
except:
loss_tot_dict[k] = v
else:
try:
loss_tot_dict[k] += v.item()
except:
loss_tot_dict[k] += v
start_idx, end_idx = end_idx, end_idx + B
if pred.size(0) != all_predictions[start_idx:end_idx].size(0):
pred = pred.view(labels.size(0), -1)
all_predictions[start_idx:end_idx] = pred.data.cpu()
all_targets[start_idx:end_idx] = labels.data.cpu()
if concept_uncertainty is not None:
all_certainties[start_idx:end_idx] = concept_uncertainty.data.cpu()
if class_uncertainty is not None:
all_cls_certainties[start_idx:end_idx] = class_uncertainty.data.cpu()
batch_idx += 1
for k, v in loss_tot_dict.items():
loss_tot_dict[k] = v / batch_idx
return all_predictions[:end_idx], all_targets[:end_idx], all_certainties[:end_idx], all_cls_certainties[:end_idx], loss_tot_dict